Blackwell Deep Dive: B200, B300, GB200
Nvidia B200 is the headline this week: $4.26/hr spot on CoreWeave (8-GPU nodes, US region) marks the lowest tracked B200 price since the Blackwell rollout began. On-demand starts at $5.98/hr (RunPod, single GPU). The average across 13 providers sits at $6.84/hr, up 12.4% week-over-week — still trending up but the entry point just dropped.
The critical caveat: that $4.26/hr is spot pricing — interruptible, region-locked to CoreWeave US, and available only in 8-GPU blocks ($34.11/hr total per node). If you need guaranteed on-demand B200 right now, RunPod's $5.98/hr single-GPU option is the floor.
Nvidia B300 is the most volatile card in the index this week — up 63.3% WoW. The average per-GPU rate hit $6.12/hr, with CoreWeave spot at $4.48/hr. Only 3 providers currently tracked (CoreWeave, RunPod on-demand, RunPod community spot), so this spike partly reflects thin-market amplification rather than broad demand surge. B300 is Blackwell Ultra — 288GB HBM3e, ~50% more FP4 performance vs B200 — but it's still in early rollout. Watch this one: if it stays elevated as more providers come online, it signals real enterprise demand for the upgrade path.
GB200 NVL is a ghost in the index: only 1 provider tracked (CoreWeave, $5.25/hr per GPU on-demand, 8-GPU nodes at $42/hr total). No price movement this week. This is the NVL72 superchip — two B200 dies + Grace CPU — designed for trillion-parameter LLM inference. If you're running MoE models at scale, this is the hardware. If you need anything else, it's expensive overkill.
Full GPU Price Index — Week 22, 2026
| GPU Model | Avg $/hr | Min $/hr | Max $/hr | WoW | Trend | Providers |
|---|---|---|---|---|---|---|
| Nvidia B200 (Blackwell) | $6.84 | $4.26 | $8.60 | ↑ +12.4% | Rising | 13 |
| Nvidia B300 (Blackwell Ultra) | $6.12 | $4.48 | $6.94 | ↑ +63.3% | Rising | 3 |
| GB200 NVL (Blackwell) | $5.25 | $5.25 | $5.25 | → 0% | Stable | 1 |
| Nvidia H200 | $3.90 | $0.50 | $6.30 | ↓ -21.3% | Dropping | 14 |
| Nvidia H100 | $3.40 | $1.49 | $13.50 | ↓ -6.8% | Dropping | 73 |
| Nvidia A100 (40GB) | $2.90 | $1.10 | $5.12 | ↓ -10.7% | Dropping | 13 |
| Nvidia A100 (80GB) | $2.72 | $0.125 | $19.44 | ↓ -4.7% | Dropping | 101 |
| Nvidia L40S | $1.74 | $0.42 | $12.48 | ↓ -12.2% | Dropping | 31 |
Top Price Movers This Week
- 1 RTX 3090 ↑ +145.6% — Vast.ai spot hit $0.31/hr, up from $0.1264 last week. Crypto mining demand surge likely driving this. Treat as volatile signal, not sustained market trend.
- 2 Nvidia B300 ↑ +63.3% — Thin market amplification + early Blackwell Ultra enterprise demand.
- 3 RTX 4090 ↑ +63.5% — Vast.ai spot at $0.59/hr. Same Vast.ai spot market dynamics.
- 4 V100 ↑ +48.0% — Vast.ai spot at $0.17/hr. Legacy card demand rising as teams cost-optimize down from H100.
- 5 A40 ↑ +42.9% — Vast.ai spot at $0.11/hr. Stable inference at consumer-grade cost.
- 6 Nvidia B200 ↑ +12.1% — Vast.ai spot at $4.32/hr. CoreWeave still cheaper at $4.26.
Is B300 worth it over B200 for LLM training?
Short answer: for most teams, not yet. Here's the math.
B300 gives you 288GB HBM3e (vs B200's 192GB) and ~50% more FP4 throughput — meaningful for models that actually need that memory headroom. But B300 currently has only 3 tracked providers vs 13 for B200, meaning less price competition, worse availability, and thinner spot markets. At $6.12/hr average vs B200's $6.84 average, B300 looks cheaper — but that B200 average includes CoreWeave's $4.26 spot which is genuinely the best deal in the index. The effective floor for B300 isn't there yet.
If you're training a 700B+ MoE model and genuinely need 288GB per GPU, B300 makes sense. If you're running anything under 400B, B200 with model parallelism will get you there at a lower all-in cost with better availability. The upgrade path is coming — but the timing for most teams is Q3-Q4 2026 as more providers roll out B300 capacity.